Appendix R — Data Card — CFPB Consumer Complaint Database

Provider & what it is

The Consumer Financial Protection Bureau (CFPB) publishes the Consumer Complaint Database: a public record of complaints that consumers have filed against financial companies — credit cards, mortgages, student loans, debt collection, credit reporting, bank accounts, “buy now pay later,” and more. When a person complains to the CFPB about a financial product, the complaint is routed to the company for a response; the CFPB then publishes a de-identified version of the record, including the product and sub-product, the issue the consumer raised, the company named, the consumer’s state/ZIP prefix, the company’s response, and — when the consumer opts in — a free-text narrative describing what happened.

The reveal here is twofold. First, it is a rare public dataset where the outcome variable is text: the narrative field is a natural training and evaluation set for the text-as-data and LLM-labeling work of Week 6, because a human can read a complaint and a model can try to classify it. Second, it is complaints, not incidence — it measures who complained, which is a selected, voluntary signal, not the true rate of financial harm. Treating complaint counts as if they were harm rates is the trap the whole card exists to warn against.

Coverage

Time: complaints from December 2011 to present, updated roughly daily as new complaints clear the publication process. [CHECK] exact update cadence for your pinned vintage.
Size: millions of complaints cumulatively; hundreds of thousands per recent year. Large but laptop-tractable when filtered by product/date — not a multi-GB-per-year file like HMDA.
Narratives: only a subset of complaints include a published free-text narrative — the consumer must opt in to publication, and the CFPB scrubs personal identifiers. So the text-analysis sample is a selected slice of the already-selected complaint population. [CHECK] current share of complaints with published narratives.
Products: the product taxonomy has changed over time (categories added, renamed, split — e.g., the credit-reporting and BNPL categories evolved), so a long time series needs the taxonomy reconciled.

Key identifiers

Complaint ID — a unique integer per complaint, the primary key.
Company — the named financial company (a string; the same firm can appear under spelling/affiliate variants, so company-level aggregation needs cleaning).
Product / Sub-product / Issue / Sub-issue — the categorical labels, the natural grouping dimensions.
State and ZIP code (3-digit prefix only, for privacy) — coarse geography.
Date received and Date sent to company — timing fields.
There is no person identifier and no household linkage; narratives are de-identified.

Access path

Public API, base https://www.consumerfinance.gov/data-research/consumer-complaints/search/api/v1/. JSON, no key required, supports filtering by date range, product, company, state, and a full-text search over narratives, with paging.
Bulk download of the full CSV/JSON is also offered from the same data-research page — fine to pull filtered, but pin the download date.

import requests, pandas as pd
BASE = "https://www.consumerfinance.gov/data-research/consumer-complaints/search/api/v1/"
# Student-loan complaints in 2024 that include a published narrative (public, no key):
params = {"product": "Student loan", "date_received_min": "2024-01-01",
          "date_received_max": "2024-12-31", "has_narrative": "true",
          "size": 100, "no_aggs": "true"}
r = requests.get(BASE, params=params, timeout=60); r.raise_for_status()
hits = r.json()["hits"]["hits"]           # each record under _source
df = pd.json_normalize([h["_source"] for h in hits])

[CHECK] exact parameter names (date_received_min, has_narrative, size) against the live API docs for your pinned vintage — the endpoint is Elasticsearch-backed and its parameter surface has shifted across releases. No secrets; the API is fully open.

License & note

Free and public, published as U.S. government open data. Cite the CFPB Consumer Complaint Database and your pinned download/access date. The CFPB’s own published caveat is the important one to repeat to students: the Bureau does not verify all the facts a consumer alleges, and publication is not a finding that a company did anything wrong. A complaint is an allegation, routed and answered, not an adjudicated fact. Quote complaint patterns as “what consumers reported,” never as proven misconduct.

Gotchas

Complaints measure complaining, not harm. Counts reflect who knew to file, had time to file, and chose to — a large, salient firm with many customers will rack up complaints without being worse per-customer. Normalize by something (accounts, deposits, loans) before comparing companies, and even then be careful.
Allegations, not verified facts. The CFPB does not confirm the underlying claims. Do not present complaint volume as evidence of wrongdoing.
Narrative selection on selection. Published narratives require consumer opt-in and CFPB scrubbing, on top of the already-voluntary act of complaining. Any text model trained on narratives generalizes only to complaints with published narratives, not to consumers at large — a textbook external-validity caveat.
Company-name messiness. The same firm appears under affiliates and spelling variants; firm-level analysis needs a cleaning/crosswalk step.
Taxonomy drift. Product/issue categories were added and renamed over time; reconcile the taxonomy before building a long panel.
Coarse geography. Only state and 3-digit ZIP prefix — fine for state-level work, useless for tract-level fair-lending geography (use HMDA for that).

First 10 rows (illustrative schema sketch — not real values)

Complaint records (API _source fields), shape only — one row per complaint:

complaint_id	date_received	product	sub_product	issue	company	state	zip_code	company_response	has_narrative
8000001	2024-03-14	Student loan	Federal student loan servicing	Dealing with your lender or servicer	Example Servicing LLC	VA	220xx	Closed with explanation	true
8000002	2024-03-14	Credit card	General-purpose credit card	Problem with a purchase shown on your statement	Sample Bank, N.A.	TX	750xx	Closed with monetary relief	false
8000003	2024-03-15	Credit reporting	Credit reporting	Incorrect information on your report	Example Credit Bureau	CA	941xx	In progress	true
…	…	…	…	…	…	…	…	…	…

Narrative text field (only when has_narrative = true), shape only:

complaint_id	consumer_complaint_narrative
8000001	“I have called my servicer XXXX times about an income-driven repayment recalculation and …”
8000003	“There is an account on my report that is not mine. I disputed it on XX/XX/XXXX and …”
…	…

Which chapter / lab / capstone uses it

Week 6 text-as-data and AI modules. The narrative field is a natural corpus for LLM-labeling and text-classification exercises: a model proposes a label (the issue, the sentiment, whether a monetary remedy is implied), and you evaluate it against the CFPB’s own categorical labels with the precision/recall harness from the Reading Guide Pack 6 / AI Lab Manual. The selection caveats make it an honest teaching case for responsible-use and external-validity disclosure.
Maya’s student-debt and consumer-finance thread. Complaints about student-loan servicing and credit reporting are directly relevant to the household-finance examples that run through the camp, and pair with HMDA (hmda.md) as the two CFPB-published consumer-finance sources — one application-level and near-census, one complaint-level and voluntary.
A self-directed capstone extension source for a student who wants a text-heavy project outside the five canonical capstone datasets, with the standing warning that complaint counts are a selected signal, not a harm rate.

# Data Card — CFPB Consumer Complaint Database ## Provider & what it is The **Consumer Financial Protection Bureau (CFPB)** publishes the **Consumer Complaint Database**: a public record of complaints that consumers have filed against financial companies — credit cards, mortgages, student loans, debt collection, credit reporting, bank accounts, "buy now pay later," and more. When a person complains to the CFPB about a financial product, the complaint is routed to the company for a response; the CFPB then publishes a de-identified version of the record, including the **product** and **sub-product**, the **issue** the consumer raised, the **company** named, the consumer's **state/ZIP prefix**, the **company's response**, and — when the consumer opts in — a **free-text narrative** describing what happened. The reveal here is twofold. First, it is a rare public dataset where the *outcome variable is text*: the narrative field is a natural training and evaluation set for the text-as-data and LLM-labeling work of Week 6, because a human can read a complaint and a model can try to classify it. Second, it is **complaints, not incidence** — it measures who *complained*, which is a selected, voluntary signal, not the true rate of financial harm. Treating complaint counts as if they were harm rates is the trap the whole card exists to warn against. ## Coverage - **Time:** complaints from **December 2011** to present, updated roughly daily as new complaints clear the publication process. `[CHECK]` exact update cadence for your pinned vintage. - **Size:** millions of complaints cumulatively; hundreds of thousands per recent year. Large but laptop-tractable when filtered by product/date — not a multi-GB-per-year file like HMDA. - **Narratives:** only a *subset* of complaints include a published free-text narrative — the consumer must opt in to publication, and the CFPB scrubs personal identifiers. So the text-analysis sample is a selected slice of the already-selected complaint population. `[CHECK]` current share of complaints with published narratives. - **Products:** the product taxonomy has changed over time (categories added, renamed, split — e.g., the credit-reporting and BNPL categories evolved), so a long time series needs the taxonomy reconciled. ## Key identifiers - **Complaint ID** — a unique integer per complaint, the primary key. - **Company** — the named financial company (a string; the same firm can appear under spelling/affiliate variants, so company-level aggregation needs cleaning). - **Product / Sub-product / Issue / Sub-issue** — the categorical labels, the natural grouping dimensions. - **State** and **ZIP code** (3-digit prefix only, for privacy) — coarse geography. - **Date received** and **Date sent to company** — timing fields. - There is **no person identifier** and no household linkage; narratives are de-identified. ## Access path - **Public API**, base `https://www.consumerfinance.gov/data-research/consumer-complaints/search/api/v1/`. JSON, **no key required**, supports filtering by date range, product, company, state, and a full-text search over narratives, with paging. - **Bulk download** of the full CSV/JSON is also offered from the same data-research page — fine to pull filtered, but pin the download date. ```python import requests, pandas as pd BASE = "https://www.consumerfinance.gov/data-research/consumer-complaints/search/api/v1/" # Student-loan complaints in 2024 that include a published narrative (public, no key): params = {"product": "Student loan", "date_received_min": "2024-01-01", "date_received_max": "2024-12-31", "has_narrative": "true", "size": 100, "no_aggs": "true"} r = requests.get(BASE, params=params, timeout=60); r.raise_for_status() hits = r.json()["hits"]["hits"] # each record under _source df = pd.json_normalize([h["_source"] for h in hits]) ``` `[CHECK]` exact parameter names (`date_received_min`, `has_narrative`, `size`) against the live API docs for your pinned vintage — the endpoint is Elasticsearch-backed and its parameter surface has shifted across releases. No secrets; the API is fully open. ## License & note **Free and public**, published as U.S. government open data. Cite the CFPB Consumer Complaint Database and your pinned download/access date. The CFPB's own published caveat is the important one to repeat to students: the Bureau **does not verify** all the facts a consumer alleges, and publication is not a finding that a company did anything wrong. A complaint is an *allegation*, routed and answered, not an adjudicated fact. Quote complaint patterns as "what consumers reported," never as proven misconduct. ## Gotchas - **Complaints measure complaining, not harm.** Counts reflect who knew to file, had time to file, and chose to — a large, salient firm with many customers will rack up complaints without being worse per-customer. Normalize by something (accounts, deposits, loans) before comparing companies, and even then be careful. - **Allegations, not verified facts.** The CFPB does not confirm the underlying claims. Do not present complaint volume as evidence of wrongdoing. - **Narrative selection on selection.** Published narratives require consumer opt-in *and* CFPB scrubbing, on top of the already-voluntary act of complaining. Any text model trained on narratives generalizes only to *complaints with published narratives*, not to consumers at large — a textbook external-validity caveat. - **Company-name messiness.** The same firm appears under affiliates and spelling variants; firm-level analysis needs a cleaning/crosswalk step. - **Taxonomy drift.** Product/issue categories were added and renamed over time; reconcile the taxonomy before building a long panel. - **Coarse geography.** Only state and 3-digit ZIP prefix — fine for state-level work, useless for tract-level fair-lending geography (use HMDA for that). ## First 10 rows (illustrative schema sketch — not real values) *Complaint records (API `_source` fields), shape only — one row per complaint:* | complaint_id | date_received | product | sub_product | issue | company | state | zip_code | company_response | has_narrative | |---|---|---|---|---|---|---|---|---|---| | 8000001 | 2024-03-14 | Student loan | Federal student loan servicing | Dealing with your lender or servicer | Example Servicing LLC | VA | 220xx | Closed with explanation | true | | 8000002 | 2024-03-14 | Credit card | General-purpose credit card | Problem with a purchase shown on your statement | Sample Bank, N.A. | TX | 750xx | Closed with monetary relief | false | | 8000003 | 2024-03-15 | Credit reporting | Credit reporting | Incorrect information on your report | Example Credit Bureau | CA | 941xx | In progress | true | | … | … | … | … | … | … | … | … | … | … | *Narrative text field (only when `has_narrative = true`), shape only:* | complaint_id | consumer_complaint_narrative | |---|---| | 8000001 | "I have called my servicer XXXX times about an income-driven repayment recalculation and …" | | 8000003 | "There is an account on my report that is not mine. I disputed it on XX/XX/XXXX and …" | | … | … | ## Which chapter / lab / capstone uses it - **Week 6 text-as-data and AI modules.** The narrative field is a natural corpus for LLM-labeling and text-classification exercises: a model proposes a label (the issue, the sentiment, whether a monetary remedy is implied), and you evaluate it against the CFPB's own categorical labels with the precision/recall harness from the Reading Guide Pack 6 / AI Lab Manual. The selection caveats make it an honest teaching case for *responsible-use and external-validity disclosure*. - **Maya's student-debt and consumer-finance thread.** Complaints about student-loan servicing and credit reporting are directly relevant to the household-finance examples that run through the camp, and pair with HMDA (`hmda.md`) as the two CFPB-published consumer-finance sources — one application-level and near-census, one complaint-level and voluntary. - A self-directed **capstone extension** source for a student who wants a text-heavy project outside the five canonical capstone datasets, with the standing warning that complaint counts are a selected signal, not a harm rate.